Evaluation of neural network ensemble approach to predict from a data stream

Abstract

We have recently worked out a method for building reliable predictive models from a data stream of real estate transactions which applies the ensembles of genetic fuzzy systems and neural networks. The method consists in building models over the chunks of a data stream determined by a sliding time window and enlarging gradually an ensemble by models generated in the course of time. The aged models are utilized to compose ensembles and their output is updated with trend functions reflecting the changes of prices in the market. In the paper we present the next series of extensive experiments to evaluate our method with the ensembles of artificial neural networks. We examine the impact of the number of aged models used to compose an ensemble on the accuracy and the influence of the degree of polynomial trend functions employed to modify the results on the performance of neural network ensembles. The experimental results were analysed using statistical approach embracing nonparametric tests followed by post-hoc procedures designed for multiple N×N comparisons.